RESEARCH ARTICLE Open Access

Understanding personalized dynamics toinform precision medicine: a dynamic timewarp analysis of 255 depressed inpatientsK. Hebbrecht1,2*, M. Stuivenga1,2, T. Birkenhäger1,2,3, M. Morrens1,2, E. I. Fried5, B. Sabbe1,2 and E. J. Giltay1,2,4*

Abstract

Background: Major depressive disorder (MDD) shows large heterogeneity of symptoms between patients, butwithin patients, particular symptom clusters may show similar trajectories. While symptom clusters and networkshave mostly been studied using cross-sectional designs, temporal dynamics of symptoms within patients may yieldinformation that facilitates personalized medicine. Here, we aim to cluster depressive symptom dynamics throughdynamic time warping (DTW) analysis.

Methods: The 17-item Hamilton Rating Scale for Depression (HRSD-17) was administered every 2 weeks for amedian of 11 weeks in 255 depressed inpatients. The DTW analysis modeled the temporal dynamics of each pair ofindividual HRSD-17 items within each patient (i.e., 69,360 calculated “DTW distances”). Subsequently, hierarchicalclustering and network models were estimated based on similarities in symptom dynamics both within eachpatient and at the group level.

Results: The sample had a mean age of 51 (SD 15.4), and 64.7% were female. Clusters and networks based onsymptom dynamics markedly differed across patients. At the group level, five dynamic symptom clusters emerged,which differed from a previously published cross-sectional network. Patients who showed treatment response orremission had the shortest average DTW distance, indicating denser networks with more synchronous symptomtrajectories.

Conclusions: Symptom dynamics over time can be clustered and visualized using DTW. DTW represents a promisingnew approach for studying symptom dynamics with the potential to facilitate personalized psychiatric care.

Keywords: Major depressive disorder, Routine outcome monitoring, Cluster analysis, Symptom dynamics, Inter-individual variation, Intra-individual variation, Symptom trajectories

BackgroundDepression is defined by its symptoms (such as a sadmood and insomnia) that are correlated with each other.The dominant explanation in the field has been thatthese relations stem from a shared causal origin, a

perspective termed the common cause framework [1, 2].The contemporary conceptualization for major depres-sive disorder (MDD) is similar to that of other medicalconditions in that it assumes all observable depressivesymptoms are caused by an underlying disease construct[1, 3]. In research, symptoms are usually added up tosum scores, and thresholds are used to indicate case sta-tus. This approach assumes that symptoms are equiva-lent, causally independent, and roughly interchangeableindicators of the underlying disease construct [4]. This

© The Author(s). 2020 Open Access This article is licensed under a Creative Commons Attribution 4.0 International License,which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you giveappropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate ifchanges were made. The images or other third party material in this article are included in the article's Creative Commonslicence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commonslicence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtainpermission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/.The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to thedata made available in this article, unless otherwise stated in a credit line to the data.

* Correspondence: kaat.hebbrecht@emmaus.be; E.J.Giltay@lumc.nl1Collaborative Antwerp Psychiatric Research Institute (CAPRI), Department ofBiomedical Sciences, University of Antwerp, Stationsstraat 22c, 2570 Duffel,BelgiumFull list of author information is available at the end of the article

Hebbrecht et al. BMC Medicine (2020) 18:400 https://doi.org/10.1186/s12916-020-01867-5

conceptual framework has dominated depression re-search over the past decades: the inclusion criteria in re-search studies were based on the syndromal DSMdiagnoses of MDD, and the unweighted sum scores ofdepression rating scales were used as a measure for se-verity and treatment response [5] (e.g., Hamilton RatingScale for Depression [HRSD] [6] and Montgomery-Åsberg Depression Rating Scale [MADRS]) [7]. How-ever, years of research have shown slow progress in thesearch for the underlying risk factors and biomarkers ofthe unitary construct MDD [4]. The need for a new, sci-entifically sound approach for conceptualizing depres-sion is warranted.Increasing evidence points towards the multidimen-

sional character of MDD with a high degree of symp-tomatic variability between and within patients [4, 8].Individual symptoms are mutually interacting and caus-ing each other [9], and they have different risk factors[10, 11], underlying biology [11–13], psychosocial im-pact [14], and course trajectories [5]. Recent years havetherefore seen a shift in the conceptualization of de-pression towards a network perspective where the de-pressive syndrome is hypothesized to stem from mutualcausal relations among components of the system, suchas depression symptoms [9, 15]. Furthermore, patientsmanifest specific depression symptom profiles withpreferential responses to different treatments. Conse-quently, there is increasing recognition of the import-ance of investigating individual symptoms and theirtimely evolution, both within individual patients and ingroups of patients [16, 17]. This is also in line with theaims of the Research Domain Criteria (RDoC) projectto deconstruct psychiatric disorders by analyzing thedynamics (e.g., symptom trajectories over time) that lieat their basis [18, 19].Several factor analytic studies of the HRSD-17 have

tried to tackle the symptomatic diversity of MDD bymeans of identifying homogeneous symptom groupswithin MDD. Although there was evidence for a “generaldepression” and “insomnia” factor, the overall resultswere inconsistent, with reported factors ranging fromtwo to eight [20–23]. Furthermore, factors seemed tochange over time [24] and were poorly generalizable toother populations. Hierarchical cluster analysis is anotherstatistical method to decompose MDD into homogeneoussymptom groups, and comparable results (“general depres-sion,” “insomnia”) have been found with this approach [25,26]. Network analysis is a more recent approach thatexpands further on studying the symptom correlations byinvestigating the influence of symptoms on each other [9].Both factor and network analyses were mostly conductedon cross-sectional data, and consequently, they did nottake the temporal dynamics of symptoms into account.Furthermore, both techniques almost exclusively studied

aggregated patient data without studying the intra-individual symptom heterogeneity.Routine outcome monitoring (ROM) entails the col-

lection of clinical data at baseline and at regular time in-tervals thereafter in order to monitor disease severity aswell as the clinical course during treatment. ROM mayprovide feedback to both the clinician and the patientand enable “patient-centered research” [27, 28]. Time-series ROM data enable the capturing of dynamics ofsymptoms over time using dynamic time warping (DTW).DTW is a widely used statistical algorithm [29, 30],though not yet in psychological and psychiatric research.It is an effective clustering strategy for time-series dataacross a broad range of application domains [31]. Exam-ples of biomedical applications are speech recognition[16], gait pathology [32–34], and electro-cardiogram ana-lysis [35]. The DTW approach could be well-suited tocluster individual symptoms based on the temporal fea-tures that they share, using ROM or ecological moment-ary assessment EMA [36] time-series data.In this study, we utilize depression symptom data from

a clinical ROM regime, every 2 weeks, of 255 depressedinpatients and present the first implementation of DTWtime-series analysis on depression symptom trajectories.This paper is built upon a dual structure in which theDTW analysis is introduced both for intra-individual(i.e., idiographic) and inter-individual (i.e., nomothetic)analysis [37]. In the idiographic analysis, we aim to as-sess the dynamics and covariation of changes in symp-toms over time within each individual patient with twoor more assessments and to estimate the symptom clus-ters and networks within each patient. In the nomotheticanalysis, we aim to study the aggregated dynamics of in-dividual symptoms to yield systematic patterns acrosspatients.

MethodsSample and settingFrom the original study sample of 276 consecutive pa-tients (i.e., included in the cohort study in the order thatthey were admitted), 21 patients had only one HRSD-17measurement due to a short hospitalization period or re-fusal to participate, yielding 255 (91.6%) patients in-cluded in the current analysis. Thus, we included 255adult patients consecutively admitted to a tertiary psy-chiatric hospital in Duffel, Belgium, and fulfilling theMINI-Plus diagnosis, based on the DSM-IV criteria, of adepressive episode as part of an MDD or bipolar dis-order (BD). In order to obtain a representative sample ofdepressed participants, exclusion criteria were minimal.We did not include patients with (comorbid MINI-Plus)psychotic disorders (including schizoaffective disorder)or with a dependency on alcohol or drugs within 12months prior to hospitalization. Moreover, patients with

Hebbrecht et al. BMC Medicine (2020) 18:400 Page 2 of 15

insufficient mastery of the Dutch language were notincluded.

TreatmentInpatients received treatment as usual which was basedon evidence-based guidelines and consisted of pharma-cotherapy, (group) psychotherapy, or a combination ofboth. These guidelines for diagnosis and treatment wereformulated by the Dutch Association of Psychiatry, oftenin association with the associations of psychology andgeneral practitioners (www.trimbos.nl, www.nvvp.net).Psychotropic medication at baseline was coded into fivedichotomous variables: antidepressants, mood stabilizers,antipsychotics, benzodiazepines, and stimulants. Re-sponse was defined as a ≥ 50% reduction of the HRSD-17 compared to the baseline assessment. Remission wasdefined as scoring ≤ 7 on the HRSD-17.

MeasurementsThe present study was part of a larger follow-up studyinvestigating the feasibility of ROM in the UniversityPsychiatric Centre in Duffel [38]. ROM were done atbaseline and every 2 weeks thereafter during the clinicaladmission which lasted from 2 weeks to 16months.ROM consisted of a test battery assessing overall mentalwell-being, quality of life, and mood (including theHamilton Depression Rating Scale-17). The data pre-sented in this article represent the collected data fromthe period April 2015 through February 2018.The HRSD-17 consists of 17 items on a Likert scale,

ranging from either 0 to 4 (for 9 items) or 0 to 2 (for 8items). The internal reliability of the HRSD-17 is ad-equate with most studies reporting a Cronbach’s alphaof ≥ 0.70. It has a good retest and interrater reliability(above 0.80) when assessed over an interval rangingfrom 1 to 30 days [21]. The Omega and Cronbach’salpha in our sample of 255 patients at baseline were only0.49 and 0.52, respectively. The Cronbach’s alpha im-proved over time with a score of 0.74 after 2, 0.77 after4, and 0.79 after 6 weeks. The total score ranges from 0to 52, and higher scores indicate greater severity, but inthe present study, we focus on the trajectories of the 17individual items only. In order to improve interrater reli-ability, Hamilton Depression Rating Scale training ses-sions were organized every 3 months among the in total6 assessors, during which video-recorded interviews withpatients were rated and discussed to reach consensus. Intotal, they conducted 1480 HRSD-17 assessments in 255patients, with an average of 5.8 assessments per patient.

Statistical analysisDTW is an approximate pattern detection algorithm thatcan measure the similarity between two time-series. Ituses a dynamic (i.e., stretching and compressing)

programming approach to minimize a predefined dis-tance measure (e.g., Euclidean distance), in order for thetwo time-series to become optimally aligned through awarping path. The “optimal” alignment minimizes thesum of distances between the aligned elements. The“dtw” (version 1.20.1), “pheatmap” (version 1.0.12), “par-allelDist” (version 0.2.4), and “qgraph” (version 1.6.2)packages for the R statistical software were used (R ver-sion 3.6.0; R Foundation for Statistical Computing,Vienna, Austria, 2016. URL: https://www.R-project.org/).The idiographic approach per patient was followed by

a nomothetic approach to study the depression symptompatterns both within individual patients and in the wholesample of 255 patients. The subsequent methodologicalsteps and statistical methods are described below.

Intra-individual approachWe first aimed to cluster individual symptoms based onthe temporal features that they share within each indi-vidual patient. The clustering of symptom trajectoriesbased on DTW consisted of two steps. First, the DTWdistance between each pair of symptom trajectories wascalculated. This is illustrated in Fig. 1 with the exampleof two HRSD-17 item time-series (item 1 “depressedmood” and item 7 “work and interest”) of a single pa-tient. The temporal scoring (per 2 weeks) on the givenitems is seen in Fig. 1a, with the two items for which theDTW distance is calculated shown in red. This patienthad 14 assessments over a period of 26 weeks. The tra-jectories of items 1 “depressed mood” and 7 “work andinterest” over time are plotted in Fig. 1b. The deforma-tions of the time axes between both items are added,which brings the two time-series as close as possible toeach other, in which all elements must be matched.Next, the calculation of the shortest path between thetwo time-series is shown in Fig. 1c. The two time-serieswere aligned in time with compressions and expansions.The “symmetricP0” step pattern was used as the dy-namic time warping algorithm to match the two se-quences, resulting in the red “warping path.” A Sakoe-Chiba Band of 2 was used in order for the severity scoresto be matched to a maximum of plus or minus two timepoints (plus or minus a maximum of 4 weeks). Resultingfrom the DTW method, a distance measure (d) is pro-duced: items with the best alignment, having a moresimilar slope and other dynamics (i.e., changes that co-vary over time), resulted in the smallest distance. Thedistance measures of each of the 17 time-series of indi-vidual HRSD-17 items are grouped in a distance matrix,comprising (172 − 17)/2 = 136 distances for each individ-ual patient.Second, this matrix of 136 distances was presented in

a heatmap and used in a hierarchical cluster analysis anda symptom network per patient. For the hierarchical

Hebbrecht et al. BMC Medicine (2020) 18:400 Page 3 of 15

Fig. 1 (See legend on next page.)

Hebbrecht et al. BMC Medicine (2020) 18:400 Page 4 of 15

cluster analysis, each item is initially assigned to its owncluster, and then the algorithm proceeds iteratively, ateach stage joining the two most similar clusters, con-tinuing until there is just a single cluster. We assumed 3clusters for each patient, for illustrative purposes only,to enable easier recognition of the symptom with themore similar trajectories. We excluded all symptomswith a score of 0 throughout follow-up, as these tendedto cluster together most strongly as these symptom pairswill have a distance of 0. At each stage, distances be-tween clusters are recomputed by the Lance-Williamsdissimilarity update formula according to the “Ward.D2”clustering methods. With “Ward.D2,” the total within-cluster variance is minimized, and the dissimilarities aresquared before cluster updating.Using the “qgraph” package, the structure of the net-

work based on the distance matrix was visualized per pa-tient, providing another way of graphical presentation ofthe clusters. We followed the recommendations on net-work analysis written by the developers of the R package[39]. A network with up to 17 nodes (representing theindividual HRSD-17 depression symptoms) is obtainedand, connecting them, the edges representing the dis-tances between symptom trajectories. The thickness ofthe edges indicates the strength of the longitudinal elas-tic covariation (thicker edges represent a shorter dis-tance between the two symptom trajectories).

Inter-individual approachNext, we aimed to study the aggregated dynamics of in-dividual symptoms to yield systematic patterns over timeacross patients. In this second part, we aimed to build ageneralizable hierarchy of symptom clusters based ontheir shared temporal features. First, the 136 distanceswere averaged over the 255 patients, weighted for thenumber of assessments that were done for each of thepatients (ranging from 2 through 17). Second, thismatrix of 255 mean distances was used for thegeneralizable hierarchical cluster analysis. A scree plotwas constructed displaying the heights in a downwardcurve and the elbow rule (i.e., the point where the graphleveled off) was used to determine the most appropriatenumber of clusters.The “Distatis” algorithm from the “DistatisR” package

was used to check whether using the actual 255 distancematrices instead of one mean distance matrix yieldedsimilar clusters. Distatis is a generalization of classicalmultidimensional scaling (MDS), based on a three-way

principal component analysis, to analyze a set of distancematrices. In order to compare these distance matrices, itcombines them into a common structure called a com-promise and then projects the original distance matricesonto this compromise. Compromise factors are calcu-lated and plotted in the compromise space, with eachcomponent been given the length corresponding to itseigenvalues. We plotted each of the 17 HRSD symptomson an X-Y plane according to their first and secondcompromise factor values.In the following step, we investigated two centrality

metrics, being closeness centrality and degree centrality[40] for the average distance matrix. Degree centrality isbased on the number and strengths of connections eachsymptom has. Closeness centrality also takes the globalnetwork structure into account because it measures theaverage distance of a certain symptom to all other symp-toms. Applied on the DTW data, closeness is inverselyproportional to the mean DTW to all other symptomsand, in this way, indicates which symptom trajectory isthe most similar to that of other symptoms.Finally, we computed the average DTW distance among

all symptom trajectories for each patient. Symptoms thatscored consistently zero were deleted from these analysesfor that particular patient, as all such symptoms would re-sult in distances of zero. Shorter average DTW distancesreflected denser interconnections between symptoms, andlonger average DTW distances reflected looser longitu-dinal connectivity between symptoms. In order to investi-gate the relationship between network density andreaching response and remission, we calculated the resid-uals of the regression with the number of assessments andthe HDRS sum score. These residuals were plotted usingbox plots according to whether response and remissionwere reached, and we performed Wilcoxon signed-ranktests to compare the two samples.The analyses used the packages “dtw” (version 1.20.1),

“pheatmap” (version 1.0.12), “parallelDist” (version0.2.4), “qgraph” (version 1.6.2), and “DistatisR” (version)for the R statistical software (R version 3.6.0; R Founda-tion for Statistical Computing, Vienna, Austria, 2016). Asample code (with data from the of 2 exemplar patientsof Fig. 2) can be found in Additional file 1.

ResultsPatient characteristicsTable 1 shows the demographic and clinical characteris-tics and the use of psychotropic medication of the

(See figure on previous page.)Fig. 1 For a single patient, the individual HRSD-17 item scores over time are shown (a). The DTW method uses a dynamic (i.e., stretching andcompressing) programming approach to minimize a predefined distance measure (e.g., Euclidean distance), in order for the two time-series tobecome optimally aligned through a warping path (b). The optimal warping route between items 1 and 7 is shown (c). Using the “symmetricP0”step pattern and a Sakoe-Chiba Band of 2, this yields a final DTW distance of 13 (d)

Hebbrecht et al. BMC Medicine (2020) 18:400 Page 5 of 15

Fig. 2 (See legend on next page.)

Hebbrecht et al. BMC Medicine (2020) 18:400 Page 6 of 15

included patients. Patients had a mean age of 50.9 years(standard deviation [SD] = 15.4), and 165 were women(64.7%). A bipolar disorder was diagnosed in 48 patients(18.8%). The mean duration of illness was 11.2 ± 15years. For 56 patients (22%), the current episode was thefirst depressive episode. The baseline HRSD-17 scorewas 20.7 (SD 4.6) on average, and 79.6% of the patientsused antidepressants. Of the 255 patients, 169 showedtreatment response and 128 remission at the end of ad-mission. The median duration of hospitalization was 11weeks, and the total number of assessments was 1480,with a mean of 5.8 and a median of 5 HRSD-17 assess-ments per patient.

Intra-individual approachIn Fig. 2, the DTW analyses of 2 exemplar patients areshown. We will discuss these two exemplar patients inorder to demonstrate the opportunities of the DTWclustering method to inform clinical practice. By com-paring the results from patients 196 and 201, we canalready see a high degree of inter-individual variability insymptom trajectories.Patient 196 was a 55-year-old female presenting with

psychotic depression. At admission, anhedonia, insom-nia, and psychic anxiety were overtly present. The anx-ious preoccupations disabled her in engaging anypsychotherapy program at the start of hospitalization. Atreatment with electroconvulsive therapy (ECT) led to aresolution of the most central symptoms (symptom withmost dense connections with other symptoms, e.g., de-pressed mood, feelings of guilt, and somatic anxiety)during the hospitalization of 2 months. Although sheremained to score relatively high on the HRSD-17 symp-toms “work and interests,” “psychic anxiety,” and“insight,” she could be discharged after the resolution ofthe majority of her depressive symptoms. The central(red) symptoms tended to fluctuate most strongly to-gether over time. Furthermore, due to the presence ofresidual symptoms that were resistant to ECT treatment,we formulated an advice for ambulatory psychologicaltherapy to focus on these persistent (blue) symptoms ofinsight, engagement in activities, and psychic anxiety asa cornerstone of further treatment.Patient 201 was a 38-year-old female who presented

with a severely depressed mood and suicidal thoughts.She described her depressive complaints as an overpow-ering sense of feeling down and agitated. There was no

loss of appetite or weight loss. A treatment with nortrip-tyline and trazodone (for her sleeping problems, mainlymiddle insomnia) was started. The sleeping problemsimproved quickly. Her depressed mood and suicidalthoughts did not change at the beginning of treatment.Treatment with lithium, because of a suspicion of under-lying bipolar disorder, led to a quick resolution of the moodand anxiety symptoms. Loss of sexual interest was initiallynot present but commenced during hospitalization, possiblyas a side effect of treatment.

Inter-individual approachFigure 3 shows the nomothetic analysis of the 255 pa-tients. A total of five clusters emerged, based on theelbow method in the scree plot (see Fig. 3a). The hier-archical cluster analysis was estimated based on theaverage weighted distance matrix (Fig. 3b). These clus-ters consisted of symptoms with a similar course trajec-tory: (1) core symptoms (2 items: “depressed mood,”“work and interests”), (2) sleep symptoms (3 items: late,middle, and early insomnia), (3) distress (2 items: “guilt,”“psychic anxiety”), (4) somatic symptoms (2 items: “geni-tal symptoms,” “general somatic symptoms”), and (5)inner turmoil (8 items: insight, weight loss, hypochon-driasis, gastro-intestinal symptoms, somatic anxiety, agi-tation, retardation, and suicide). As is shown inAdditional file 2: Fig. S1, the network plots did notchange significantly when excluding all symptoms with ascore of 0 throughout follow-up.In the following step, we analyzed the actual 255 dis-

tance matrices, instead of one mean distance matrix,using Distatis. Figure 3c shows the Distatis compromiseplot in which each HRSD-17 item is plotted accordingto their first and second compromise factor values. Thedistribution pattern of the HRSD-17 items in the com-promise plot shows a comparable pattern to the ob-tained hierarchical clusters, corroborating the obtainedfive clusters. The clusters corroborated those found withthe hierarchical cluster analysis on the average distancematrix. Next, the average distance matrix was visuallypresented in a network graph in Fig. 3d.Figure 4 shows the two centrality measures based on

the network from the average distance matrix. Thesecan inform us on which symptoms globally tend to co-vary together over time. The items from the “inner tur-moil” show the highest degree centrality and closenesscentrality scores, indicating that they covaried most

(See figure on previous page.)Fig. 2 DTW analysis of HRSD-17 symptoms for patient no. 196 and patient no. 201 (a). Heatmap (symptoms that show high correlation are givena “hot” red color, and those that are not correlated are given a “cold” blue color) (b). Dendrogram based on the clustering of DTW distances of15 of the non-zero HRSD-17 item scores over time (c). Network graph based on the distance matrix: connections between symptoms (edges)indicate distances between symptom trajectories (d). Centrality statistics of the network graph: centrality is based on the number and strengthsof connections each symptom has. Closeness also takes the global network structure into account (e)

Hebbrecht et al. BMC Medicine (2020) 18:400 Page 7 of 15

strongly with other HRSD-17 items. Items constitutingthe “insomnia” or “somatic symptom” cluster showedlower centrality which suggests that these symptoms be-haved in a more independent manner over time com-pared to the other HRSD-17 items.The evolution of the mean HRSD-17 item levels over

time are visualized in Fig. 5a using mixed models peritem. The eight items with a range from 0 to 2 (three in-somnia items: gastro-intestinal complaints, general som-atic and genital symptoms, insight and weight loss) werescaled to a range from 0 to 4 in order to make a com-parability between all items possible. The HRSD-17items “depressed mood,” “work and interests,” “generalsomatic,” and “genital symptoms” had the highest base-line severity. The items “insight” and “weight loss” hadthe lowest mean scores and stayed relatively low duringhospitalization. Figure 5b shows the intercepts andslopes of the 17 mixed models for the individual longitu-dinal trajectories. The intercepts indicated that genital,general somatic, and depressed mood symptoms gener-ally scored the highest at baseline. The slopes of the lin-ear model revealed that depressed mood showed thesteepest decline over time.As shown in Fig. 6, patients that reached response or

remission during hospitalization had significantly shorteraverage distances among symptoms than patients whofailed to reach response or remission. That is, patientsreaching response or remission mostly had on average amore densely connected symptom network (based onthe mean DTW analysis). We excluded symptoms thatscored zero throughout the admission, yet when thesesymptoms were included, this resulted in similar findings(see Additional file 2: Fig. S2A). In addition, not adjust-ing for HRSD-17 total scores at baseline did not alterthe results (see Additional file 2: Fig. S2B). Exploring thedifference in network connectivity between unipolar andbipolar depressed patients revealed a denser symptomnetwork in bipolar than in unipolar patients (see Add-itional file 2: Fig. S3).

DiscussionThe present study is the first to analyze the time-seriesof depression symptoms using DTW analyses in psychi-atric inpatients. We applied the DTW computationalmethod to estimate and visualize similarities in symptomtrajectories and to yield clusters of symptoms with simi-lar course trajectories both at the patient level and at thegroup level. Both the intra- and inter-individual analysesmay help to increase our insight into the dynamicalcomplexity of symptom trajectories in severely depressedinpatients. Furthermore, combining ROM techniqueswith automated feedback for the clinician based on themethods—as introduced here—proved useful to inform

Table 1 Characteristics and medication use of 255 consecutivedepressed inpatients

Variable Mean (SD)/no. (%)

Demographic characteristics

Female (%) 165 (64.7)

Bipolar disorder 48 (18.8)

Age in years, mean (SD) 50.9 ± 15.4

Educationa,b (n (%))

- Lower 41 (16.1)

- Intermediate 119 (46.7)

- Higher 93 (36.5)

Work status (n (%))

- Unemployed 149 (58.4)

- Employed 93 (36.5)

- Others (student/voluntary service) 13 (5.1)

Marital status (n (%))

- Married 109 (42.7)

- Divorced/widowed 65 (25.5)

- Never married 81 (31.8)

Living situationc (n (%))

- Living alone 76 (29.8)

- Living with partner 96 (37.6)

- Living with family 83 (32.5)

Clinical characteristics

Index (first) depressive episode (n (%)) 56 (22)

History of 4 or more depressive episodes (n (%)) 76 (29.8)

Lifetime substance abuse/dependencyd

- Alcohol 27 (10.6)

- Drugs (THC, hard drugs, benzodiazepines) 5 (2.0)

Melancholic features 159 (62.4)

ROM baseline total scores

- HRSD-17 20.7 ± 4.6

- BDI-II 33.7 ± 9.2

Baseline medication use

- Antidepressants 203 (79.6)

- Antipsychotics 118 (46.3)

- Mood stabilizers 34 (13.3)

Responders (%) 169 (66.3)

Remitters (%) 128 (50.2)

Data are mean (SD) or no. (%), when appropriateROM routine outcome monitoring, HRSD-17 17-item Hamilton Rating Scale forDepression, BDI-II Beck Depression InventoryaLower education, general basic education only; intermediate education,middle vocational education; and higher education, higher vocationaleducation or universitybTwo missing values for educationcLiving alone includes living in a home for the elderly and the conventdInvestigated using the MINI modules on substance abuse and suicidality

Hebbrecht et al. BMC Medicine (2020) 18:400 Page 8 of 15

and facilitate clinical decision making. Overall, threemajor findings are worth discussing in more detail.We first focused on the individual symptom dynamics

that proved to be highly variable across individuals andthus idiosyncratic [41]. This finding supports previousconcerns on the use of sum scores for assessing treat-ment outcome, since sum scores do not represent thisdynamical symptom complexity well [4, 16], with a lossof substantial information that may be of clinical

relevance. The intra-individual dynamic symptom clus-ters and symptom networks, in which the edges betweensymptoms represented the dynamical relation betweenthem, allowed us to gain insight into the relative import-ance of certain symptoms for individual patients, butalso at the group level [40]. Such symptoms may causeother symptoms, which may be different for other pa-tients [9]. It could be hypothesized that targeting treat-ment on such central symptoms early in therapy may

Fig. 3 Nomothetic analyses based on all distance matrices from 255 depressed inpatients (see accompanying PDF). The scree plot displays theeigenvalues in a downward curve. The number of factors was determined using the elbow method (i.e., the point where the slope of the curve isleveling of; in our example, this is 5: after this point, the slope of the curve is nearly stable) (a). Ward’s (D2, i.e., general agglomerative hierarchicalclustering procedure) clustering criterion on the weighted mean distance matrix from 255 patients (b). Distatis analysis: the PCA of thecompromise matrix (i.e., weighted average of individual cross-product matrices) gives the position of the objects in the compromise space (c).Overview of the networks of HRSD-17 items for 255 patients (d)

Hebbrecht et al. BMC Medicine (2020) 18:400 Page 9 of 15

lead to a more rapid resolution of closely connected de-pressive symptoms [42, 43].Overall, the study of intra-individual temporal dynam-

ics of depression symptoms is rare in the literature. Agrowing field of research has focused on the develop-ment of individual dynamic networks of symptoms inwhich time-series or experience sampling methods(ESM) data are used to study the within-patient dynam-ical structure of symptoms [43–47]. These networks aremostly estimated using vector autoregression (VAR)which estimates both lagged (i.e., time minus one tem-poral) and contemporaneous (i.e., simultaneous) rela-tionships among multiple symptoms [45]. The DTWapproach represents a less constraint analysis of individ-ual symptom dynamics since the DTW distance measureaccounts for a longer time period when measuring thesimilarity between each pair of depressive symptoms (2time points). Furthermore, it provides an accessible andeasily interpretable method that can be a useful tool forclinicians and researchers for the early detection of cen-tral symptoms and the directed tailoring of treatment to-wards these symptoms. Moreover, it does not necessitatethe time-consuming ESM data collection, which can be

challenging in daily clinical practice due to the consider-able burden it puts on participants.Secondly, we focused on the group-level analyses,

which yielded five symptom clusters with more similardynamics over time across the total sample. Comparedto cross-sectional factor analytic studies, which investi-gate the co-occurrence of depressive symptoms at a cer-tain time point, we similarly found that the three sleepitems appeared consistently in one factor [20, 21]. Previ-ously, there was some support for the presence of a“general depression,” with depressed mood, guilt, sui-cide, work and interests, and psychic anxiety appearingon one factor [20–22]. We, however, found that only“depressed mood” and “work and interest” showed themost consistent trajectories over time, which repre-sented the core symptoms of depression. Somatic symp-toms did not appear on the same factor as described forcross-sectional factor analyses (“somatic symptom” or“somatized depression” consisting of somatic symptoms,weight loss/gastro-intestinal symptoms, loss of libido/genital symptoms) [20, 22]. Moreover, previous studiesfound evidence for limited longitudinal invariance,where the number of factors did not hold across time

Fig. 4 Centrality measures. The closeness centrality is the inverse of the average length of the shortest path between the focal node and everyother node in the network (i.e., the more central a node is, the closer it is to all other nodes). Degree centrality represents the connectivity, basedon the number and strengths of edges connected to it

Hebbrecht et al. BMC Medicine (2020) 18:400 Page 10 of 15

Fig. 5 Forest plot of the 17 HRSD items of two mean levels of indicators of individual longitudinal trajectories. The mixed model intercept (a)indicates which symptoms scored the highest at baseline (i.e., genital, general somatic symptoms, and depressed mood scored the highest atbaseline). The mixed model slope (b) of the linear model showed the average decline per 2-week time interval (i.e., depressed mood showed thesteepest decline over time)

Hebbrecht et al. BMC Medicine (2020) 18:400 Page 11 of 15

[24, 48] which is supported by our and previous datathat the Omegas and Cronbach’s alphas were not stableover time, but improved during hospital admission [49].An internal validation of our findings using a randomsample of 128 and 127 patients of our sample revealedthe same dynamical clusters (see Additional file 2: Fig.S4). Further validation of these findings in an independ-ent sample is necessary.Third, we found that patients who reached response

or remission during hospitalization had a more denselyconnected symptom network compared to patients thatfailed to reach response or remission. This contrastswith the findings of Van Borkulo et al. [50], who identi-fied a more densely connected cross-sectional depressionsymptom network in not remitters compared to patientsreaching remission. Although the method of quantifyingnetwork connectivity was not the same (average DTWdistance versus network comparison test), this shows,once again, the importance of studying longitudinal net-works besides cross-sectional symptom networks. Ourfindings could be related to the literature reviewed byScheffer [51], showing that networks with high connectiv-ity can change more abruptly (for better and worse) in re-sponse to external events (so-called critical transitions).

Applied on the DTW network analysis, when symptomshave a low level of connectivity, they seem to behave moreindependently from each other and in response to an ex-ternal factor such as admission and treatment, which mayhave lowered the probability of an acute response or re-mission to treatment. These findings need to be confirmedin further studies, as the definition of response and remis-sion was also based on the HRSD sum scores, which isnot independent of the DTW assessments from HRSDtime-series data.The DTW method has a promising potential for clin-

ical practice, and it builds further upon the already avail-able evidence of the value of measurement-based care inpsychiatry [52]. First, the DTW symptom clusters allowthe clinician to gain insight into the dynamics of individ-ual depression symptoms and longitudinal symptomclusters. Second, as illustrated in the two idiographicanalyses (Fig. 2), the DTW method has the potential tofacilitate clinical decision making. More specifically,treatment interventions targeted at the most centralsymptom (i.e., symptoms with the most dense connec-tions with other symptoms) could lead to a rapid reso-lution of the depressive syndrome. Third, the graphicalrepresentation of the DTW clusters is easily amenable as

Fig. 6 Average DTW distance according to response and remission. Those patients with response or remission had the shortest average distanceamong symptom trajectories, indicating denser interconnections (p by Wilcoxon signed-rank test to compare the two samples). Distances wereadjusted for the number of assessments and the baseline total HRSD-17 score

Hebbrecht et al. BMC Medicine (2020) 18:400 Page 12 of 15

a feedback tool for patients to gain more insight into thecentral symptoms that tend to covariate with a variety ofother symptoms or in symptom clusters that tend tomove in a more independent matter. This could lead toa more nuanced insight in reaching response or remis-sion or lack thereof.An important strength of our study is the use of the

innovative DTW clustering method to study the time-series of individual symptom severity scores. The DTWmethod is able to process the highly dimensional ROMtime-series data in order to reduce the complexity of thedata while still maintaining the essential characteristicsof the dataset. By using an elastic measurement, DTWprovides an optimal time alignment between two time-series. Furthermore, DTW can be accurately used insmaller datasets and individual patients [31]. Anotherstrength of our study is the relatively complete dataset ofROM data from real-world consecutive inpatients. None-theless, our results must be considered in light of some lim-itations. First, exclusively inpatients were recruited fromone center which may limit the generalizability of our re-sults to outpatients and other patient groups. Second, pa-tients were treated with a variety of different combinationsof psychotropic drugs which likely affected the course anddynamical characteristics of individual symptoms (such asconcentration difficulties in those receiving ECT). Futurestudies using data from randomized trials may help to un-ravel the influence of different treatment strategies on thedynamic symptom dimensions. Third, the HRSD-17 is notdesigned to investigate individual symptoms, and its itemsare scored on a crude scale with only three or five answercategories resulting in low variability and precision. Fourth,assessments were done with 2-week intervals, and DTWanalyses may be more useful in more frequent time-serieslike those collected with ESM. Fifth, the DTW method al-lows some flexibility in how it is applied to study MDDsymptom trajectories, e.g., in terms of the global constraint(Sakoe-Chiba Band). We adopted default settings based onsimulation studies in the prior literature and hope that fu-ture methodological studies working with psychiatric dataspecifically will investigate how robust empirical results areto changes in default settings of the DTW method, e.g.,using multiverse analyses [53].

ConclusionMDD is a heterogeneous disorder consisting of dynamicsymptom clusters that varied between patients. The useof repeated, standardized clinical rating scales yields ex-tensive information on patient-specific symptoms dy-namics. DTW may be a promising new methodology forthe study of the complex dynamic system of interactingpsychiatric symptoms [9, 15, 54] with the potential to fa-cilitate personalized psychiatry care.

Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s12916-020-01867-5.

Additional file 1. Sample R script for paper.

Additional file 2: Figure S1. Network plots based on the DTW analysesin the 255 patients. We compared the network structure based on alldata from all patients, and the analyses based only on symptoms that didnot score as 0 throughout the follow-up. Such symptom pairs with scoresof 0 will result in a distance of 0, solely because they were absent insome patients. Figure S2. Average DTW distance according to responseand remission, including symptoms that scored zero throughout thehospitalization. Figure S3. Network plots based on the DTW analyses in231 of the 255 patients, who had a clinical diagnosis of either MDD ofBD. We compared the network structure [A and B] and calculated themean distance only on symptoms that did not score as 0 throughout thefollow-up. Although the network structure was largely similar, on averagepatients with BD had a denser distance matric than patients with MDD(P=0.0028). Figure S4. Network plots of two subsamples [A and B] ofthe 255 patients. We used an automated split with a subset of 128 and127 patients, in which we conducted separate DTW analyses. Node place-ment was done by using the Procrustes algorithm (from the R Package‘networktools’), to aid the visual comparison between the two networks.As a result, configurations were brought into a similar space in which sta-tistically meaningless differences were removed without changing the fit.This analysis showed that the network (based on each of the average dis-tance matrixes were stable. The congruence coefficient was high at 0.994,when we compared both sets of compromise factors derived from eachDistatis analysis.

Abbreviationsd: Distance measure; DTW: Dynamic time warping; ECT: Electroconvulsivetreatment; HRSD-17: 17-Item Hamilton Rating Scale for Depression;MDD: Major depressive disorder; MDS: Multidimensional scaling; SD: Standarddeviation; VAR: Vector autoregression

AcknowledgementsWe thank all the clinicians who performed and coordinated the routineoutcome monitoring in the University Psychiatric Centre of Duffel.

Authors’ contributionsB.S. devised the project. B.S. and E.G. supervised the project. K.H., M.S., andE.G. processed the data. K.H. and E.G. analyzed and interpreted the patientdata with the contribution of E.F. K.H. took the lead in writing themanuscript in consultation with E.G., M.S., M.M., T.B., and E.F. All authors readand approved the final manuscript.

FundingThe study was funded by the Hercules Project (42/FA0201000/17/6780).

Availability of data and materialsThe datasets used analyzed during the current study are available from thecorresponding author on reasonable request.

Ethics approval and consent to participateOur study was approved by the Committee for Medical Ethics of theUniversity Hospital Antwerp and the Antwerp University: protocol numberB300201837075. The study complied with the Declaration of Helsinki. Datawere gathered using ROM and Good Clinical Practice. Patients wereinformed of ROM data collection, and we requested patients’ permission touse their data for research. The two patients presented in the case reportsprovided consent for publication. In order to secure patients’ confidentiality,all patient-identifiable data were removed from the database.

Competing interestsThe authors have no conflict of interest to declare.

Author details1Collaborative Antwerp Psychiatric Research Institute (CAPRI), Department ofBiomedical Sciences, University of Antwerp, Stationsstraat 22c, 2570 Duffel,

Hebbrecht et al. BMC Medicine (2020) 18:400 Page 13 of 15

Belgium. 2University Psychiatric Hospital Duffel, VZW Emmaüs, Duffel,Belgium. 3Department of Psychiatry, Erasmus Medical Center, Rotterdam, TheNetherlands. 4Department of Psychiatry, Leiden University Medical Center,Leiden, The Netherlands. 5Department of Clinical Psychology, LeidenUniversity, 2300 RA Leiden, The Netherlands.

Received: 12 July 2020 Accepted: 23 November 2020

References1. Borsboom D. Psychometric perspectives on diagnostic systems. J Clin

Psychol. 2008 Sep;64(9):1089–108.2. Schmittmann VD, Cramer AOJ, Waldorp LJ, Epskamp S, Kievit RA, Borsboom

D. Deconstructing the construct: a network perspective on psychologicalphenomena. New Ideas Psychol. 2013;31(1):43–53.

3. Borsboom D, Mellenbergh GJ, Van Heerden J. The theoretical status oflatent variables. Psychol Rev. 2003;110:203–19.

4. Fried EI. Problematic assumptions have slowed down depression research:why symptoms, not syndromes are the way forward. Front Psychol. 2015;6:309.

5. Fried EI, Nesse RM. Depression sum-scores don’t add up: why analyzingspecific depression symptoms is essential. BMC Med. 2015;13(1):72 Availablefrom: http://www.ncbi.nlm.nih.gov/pubmed/25879936. [cited 2020 Mar 15].

6. Trajković G, Starčević V, Latas M, Leštarević M, Ille T, Bukumirić Z, et al.Reliability of the Hamilton Rating Scale for Depression: a meta-analysis overa period of 49 years. Psychiatry Res. 2011;189(1):1–9.

7. Montgomery SA, Asberg M. A new depression scale designed to besensitive to change. Br J Psychiatry. 1979;134(4):382–9 Available from:https://pubmed.ncbi.nlm.nih.gov/444788/. [cited 2020 Jul 10].

8. Gibbons RD, Clark DC, Kupfer DJ. Exactly what does the Hamiltondepression rating scale measure? J Psychiatr Res. 1993;27(3):259–73.

9. Borsboom D, Cramer AOJ. Network analysis: an integrative approach to thestructure of psychopathology. Annu Rev Clin Psychol. 2013;9(1):91–121.

10. Cramer AOJ, Waldorp LJ, Van Der Maas HLJ, Borsboom D. Comorbidity: anetwork perspective. Behav Brain Sci. 2010;33:137–50.

11. Hasler G, Northoff G. Discovering imaging endophenotypes for majordepression. Mol Psychiatry. 2011;16(6):604–19 Available from: http://www.ncbi.nlm.nih.gov/pubmed/21602829. [cited 2019 Nov 8].

12. Myung W, Song J, Lim SW, Won HH, Kim S, Lee Y, et al. Genetic associationstudy of individual symptoms in depression. Psychiatry Res. 2012;198(3):400–6.

13. Kendler KS, Aggen SH, Neale MC. Evidence for multiple genetic factorsunderlying DSM-IV criteria for major depression. JAMA Psychiatry. 2013;70(6):599–607.

14. Faravelli C, Servi P, Arends JA, Strik WK. Number of symptoms,quantification, and qualification of depression. Compr Psychiatry. 37(5):307–15 Available from: http://www.ncbi.nlm.nih.gov/pubmed/8879904. [cited2020 Mar 15].

15. Kendler KS, Zachar P, Craver C. What kinds of things are psychiatricdisorders? Psychol Med. 2011;41(6):1143–50 Available from: http://www.ncbi.nlm.nih.gov/pubmed/20860872. [cited 2020 Mar 16].

16. Fried EI, Nesse RM, Zivin K, Guille C, Sen S. Depression is more than the sumscore of its parts: individual DSM symptoms have different risk factors.Psychol Med. 2014;44(10):2067–76 Available from: http://www.ncbi.nlm.nih.gov/pubmed/24289852. [cited 2020 Mar 15].

17. Beltz AM, Wright AGC, Sprague BN, PCM M. Bridging the nomothetic andidiographic approaches to the analysis of clinical data. Assessment. 2016;23(4):447–58 Available from: http://www.ncbi.nlm.nih.gov/pubmed/27165092. [cited 2020 Mar 15].

18. Insel T, Cuthbert B, Garvey M, Heinssen R, Pine DS, Quinn K, et al. ResearchDomain Criteria (RDoC): toward a new classification framework for researchon mental disorders. Am J Psychiatry. 2010;167:748–51.

19. Insel TR. The NIMH Research Domain Criteria (RDoC) project: precisionmedicine for psychiatry. Am J Psychiatry. 2014;171:395–7.

20. Shafer AB. Meta-analysis of the factor structures of four depressionquestionnaires: Beck, CES-D, Hamilton, and Zung. J Clin Psychol. 2006;62(1):123–46 Available from: http://doi.wiley.com/10.1002/jclp.20213. [cited 2020May 6].

21. Bagby RM, Ryder AG, Schuller DR, Marshall MB. The Hamilton DepressionRating Scale: has the gold standard become a lead weight? Am J Psychiatry.2004;161:2163–77.

22. Hamilton M. A rating scale for depression. J Neurol Neurosurg Psychiatry.1960;23(1):56–62 Available from: https://jnnp.bmj.com/content/23/1/56.[cited 2018 Oct 31].

23. Hamilton M. Development of a rating scale for primary depressive illness. BrJ Soc Clin Psychol. 1967;6(4):278–96.

24. Fried EI. Are more responsive depression scales really superior depressionscales? J Clin Epidemiol. 2016;77:4–6.

25. Kasper S, Dienel A. Cluster analysis of symptoms during antidepressanttreatment with Hypericum extract in mildly to moderately depressed out-patients. A meta-analysis of data from three randomized, placebo-controlledtrials. Psychopharmacology (Berl). 2002;164(3):301–8 Available from:https://pubmed.ncbi.nlm.nih.gov/12424554/. [cited 2020 Oct 15].

26. Chekroud AM, Gueorguieva R, Krumholz HM, Trivedi MH, Krystal JH,Mccarthy G. Reevaluating the efficacy and predictability of antidepressanttreatments a symptom clustering approach supplemental content. JAMAPsychiatry. 2017;74(4):370–8 Available from: https://jamanetwork.com/.

27. de Beurs E, den Hollander-Gijsman ME, van Rood YR, van der Wee NJ, GiltayEJ, van Noorden MS, van der Lem R, van Fenema E, Zitman FG. Routineoutcome monitoring in the Netherlands: practical experiences with a web-based strategy for the assessment of treatment outcome in clinical practice.Clin Psychol Psychother. 2011;18(1):1-12.

28. Washington AE, Lipstein SH. The Patient-Centered Outcomes ResearchInstitute—promoting better information, decisions, and health. N Engl JMed. 2011;365(15):e31 Available from: http://www.ncbi.nlm.nih.gov/pubmed/21992473. [cited 2019 Feb 6].

29. Sakoe H, Chiba S. Dynamic programming algorithm optimization for spokenword recognition. IEEE Trans Acoust. 1978;26(1):43–9.

30. Introduction. In: Information retrieval for music and motion. Berlin,Heidelberg: Springer Berlin Heidelberg; 2007. p. 1–13. Available from:http://link.springer.com/10.1007/978-3-540-74048-3_1. [cited 2020 Mar 30].

31. Ding H, Trajcevski G, Scheuermann P, Wang X, Keogh E. Querying andmining of time series data. Proc VLDB Endow. 2008;1(2):1542–52.

32. Dandu SR, Engelhard MM, Qureshi A, Gong J, Lach JC, Brandt-Pearce M,et al. Understanding the physiological significance of four inertial gaitfeatures in multiple sclerosis. IEEE J Biomed Heal Informatics. 2018;22(1):40–6 Available from: http://www.ncbi.nlm.nih.gov/pubmed/29300700. [cited2020 Mar 30].

33. Engelhard M, Dandu SR, Lach J, Goldman M, Patek S. Toward detection andmonitoring of gait pathology using inertial sensors under rotation, scale,and offset invariant dynamic time warping. In: Proceedings of the 10th EAIInternational Conference on Body Area Networks. ICST; 2015. Available from:http://eudl.eu/doi/10.4108/eai.28-9-2015.2261503. [cited 2020 Mar 30].

34. Li M, Tian S, Sun L, Chen X. Gait analysis for post-stroke hemiparetic patientby multi-features fusion method. Sensors (Basel). 2019;19(7):1737.

35. Zhang G, Kinsner W, Huang B. Electrocardiogram data mining based onframe classification by dynamic time warping matching. Comput MethodsBiomech Biomed Engin. 2009;12(6):701–7 Available from: http://www.ncbi.nlm.nih.gov/pubmed/19360509. [cited 2020 mar 30].

36. Csikszentmihalyi M, Larson R. Validity and reliability of the experience-sampling method. J Nerv Ment Dis. 1987;175(9):526–36.

37. Fisher AJ, Newman MG, Molenaar PCM. A quantitative method for theanalysis of nomothetic relationships between idiographic structures:dynamic patterns create attractor states for sustained posttreatmentchange. J Consult Clin Psychol. 2011;79(4):552–63.

38. Hebbrecht K, Stuivenga M, Birkenhäger T, Van Der Mast RC, Sabbe B, GiltayEJ. Symptom profile and clinical course of inpatients with unipolar versusbipolar depression. Neuropsychobiology. 2019;79:4–5 Available from:https://pubmed.ncbi.nlm.nih.gov/31655820/. [cited 2020 Jul 27].

39. Epskamp S, Borsboom D, Fried EI. Estimating psychological networks andtheir accuracy: a tutorial paper. Behav Res Methods. 2018;50(1):195–212.

40. Opsahl T, Agneessens F, Skvoretz J. Node centrality in weightednetworks: generalizing degree and shortest paths. Soc Networks. 2010;32(3):245–51.

41. Fisher AJ. Toward a dynamic model of psychological assessment:implications for personalized care. J Consult Clin Psychol. 2015;83(4):825–36.

42. Fried EI, Epskamp S, Nesse RM, Tuerlinckx F, Borsboom D. What are “good”depression symptoms? Comparing the centrality of DSM and non-DSMsymptoms of depression in a network analysis. J Affect Disord. 2016;189:314–20.

43. Epskamp S, van Borkulo CD, van der Veen DC, Servaas MN, Isvoranu AM,Riese H, et al. Personalized network modeling in psychopathology: the

Hebbrecht et al. BMC Medicine (2020) 18:400 Page 14 of 15

importance of contemporaneous and temporal connections. Clin PsycholSci. 2018;6(3):416–27 Available from: https://pubmed.ncbi.nlm.nih.gov/29805918/ [cited 2020 Jul 9].

44. Beltz AM, Wright AGC, Sprague BN, PCM M. Bridging the nomothetic andidiographic approaches to the analysis of clinical data. Assessment. 2016;23(4):447–58 Available from: http://www.ncbi.nlm.nih.gov/pubmed/27165092. [cited 2020 May 2].

45. Bringmann LF, Ferrer E, Hamaker EL, Borsboom D, Tuerlinckx F. Modelingnonstationary emotion dynamics in dyads using a time-varying vector-autoregressive model. Multivariate Behav Res. 2018;53(3):293–314.

46. Bulteel K, Tuerlinckx F, Brose A, Ceulemans E. Improved insight into andprediction of network dynamics by combining VAR and dimensionreduction. Multivariate Behav Res. 2018;53(6):853–75.

47. Fisher AJ, Reeves JW, Lawyer G, Medaglia JD, Rubel JA. Exploring theidiographic dynamics of mood and anxiety via network analysis. J AbnormPsychol. 2017;126(8):1044–56.

48. Steinmeyer EM, Möller HJ. Facet theoretic analysis of the Hamilton-D scale. JAffect Disord. 1992;25(1):53–61.

49. Fried EI, van Borkulo CD, Epskamp S, Schoevers RA, Tuerlinckx F, BorsboomD. Measuring depression over time...or not? lack of unidimensionality andlongitudinal measurement invariance in four common rating scales ofdepression. Psychol Assess. 2016;28(11):1354–67.

50. Van Borkulo C, Boschloo L, Borsboom D, Penninx BWJH, Lourens JW,Schoevers RA. Association of symptom network structure with the course oflongitudinal depression. JAMA Psychiatry. 2015;72(12):1219–26.

51. Van De Leemput IA, Wichers M, Cramer AOJ, Borsboom D, Tuerlinckx F,Kuppens P, et al. Critical slowing down as early warning for the onset andtermination of depression. Proc Natl Acad Sci U S A. 2014;111(1):87–92.

52. Trivedi MH, Rush AJ, Wisniewski SR, Nierenberg AA, Warden D, Ritz L, et al.Evaluation of outcomes with citalopram for depression using measurement-based care in STAR*D: implications for clinical practice. Am J Psychiatry.2006;163(1):28–40.

53. Steegen S, Tuerlinckx F, Gelman A, Vanpaemel W. Increasing transparencythrough a multiverse analysis. Perspect Psychol Sci. 2016;11(5):702–12.

54. Cramer AO, van Borkulo CD, Giltay EJ, van der Maas HL, Kendler KS, SchefferM, Borsboom D. Major depression as a complex dynamic system. PLoS One.2016;11(12):e0167490.

Publisher’s NoteSpringer Nature remains neutral with regard to jurisdictional claims inpublished maps and institutional affiliations.

Hebbrecht et al. BMC Medicine (2020) 18:400 Page 15 of 15